Using Vc for vectorization?

[teonnik]

Hi, great library!

I was wondering whether there have been any attempts to incorporate Vc into your code? I am aware of xsimd but it lacks the AVX512 support which is present in Vc. A follow-up, were there any issues with Boost.SIMD or Vc that prompted the development of xsimd?

[SylvainCorlay]

Quick note: avx512 and Neon support are in the roadmap for the next few months.

Xsimd and boost.simd have similarities and we did consider boost.simd but the dependency on boost was a deal breaker. Also, we have been using an older version of xsimd for a long time (other projects than xtensor).

[wolfv]

since boost simd is officially gone now (check the repo, it's more or less empty) and we just added NEON support to xsimd, I think we can close this issue.

AVX512 support will follow later this year, as well as more integer types.

[teonnik]

Well, I posted that quite a while ago and the issue is closed but I still have a few words to say.

Defining an abstraction over SIMD is a very time-consuming endeavour. I suggested Vc because the future C++ SIMD abstraction is based on it. The author of the library is the one pushing for SIMD types in C++20 http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2018/p0214r8.pdf . There are no other contenders.

Supporting various math functions like sin, cos, tan etc for SIMD types requires an enormous time investment... I am a bit worried that there is some duplication of effort that is better spent somewhere else.

[SylvainCorlay]

Supporting various math functions like sin, cos, tan etc for SIMD types requires an enormous time investment... I am a bit worried that there is some duplication of effort that is better spent somewhere else.

Yeah, this has already been done in xsimd. At the moment, we believe that xsimd is a better choice.